Emperor Penguin Colony Detection from Sentinel-2 Imagery

Noville Peninsula Case Study - January 2025

Overview

This document summarises an exploratory analysis to detect emperor penguin colonies from Sentinel-2 satellite imagery using spectral indices. Starting from a compressed RGB PNG, we progressively refined detection methods using 16-bit data and full 7-band multispectral imagery.

Key finding: Penguin colonies (guano staining) have a distinctive flat visible spectrum (R ≈ G ≈ B) compared to clean ice which slopes downward (B > G > R). This difference is captured effectively by a simple R-B spectral index.

Data

Source Format Bands Resolution
Sentinel-2 L2A GeoTIFF RGB, NIR, SWIR16, SWIR22, SCL 10m
Location Noville Peninsula, Antarctica
Date 2025-01-09

Method Evolution

Stage 1: 8-bit RGB PNG

Initial detection using simple band ordering rule:

colony_mask = (R > G) & (G > B) & (R > 100) & (R < 250)

Result: 177 pixels = 1.77 ha detected in bottom-right region

8-bit colony detection
8-bit colony detection

Lesson learned: The naive R-B index initially found false positives in shadowed areas (bottom-left). Shadows appear “warm” because blue skylight fills them less than direct illumination. The key distinction is absolute brown (R > G > B) vs relative warmth.

Stage 2: 16-bit GeoTIFF

Full dynamic range (~14,000 levels vs 255) revealed:

  • Colony region R-B anomaly: +0.004 relative to ice baseline
  • The 8-bit detection was capturing real signal, not compression noise
  • Nested confidence zones possible with finer thresholds
Confidence Threshold Area
Core R-B > 0 1.9 ha
Likely stained R-B > -0.015 11.1 ha
Influence zone R-B > -0.025 18.8 ha
16-bit multi-threshold detection
16-bit multi-threshold detection

Stage 3: 7-Band Multispectral

Added NIR (842nm), SWIR16 (1610nm), SWIR22 (2190nm) for full spectral analysis.

Spectral Signature Comparison

Spectral signature: colony vs ice
Spectral signature: colony vs ice
Band Colony Ice Difference
Blue (490nm) 0.73 0.96 -0.23
Green (560nm) 0.73 0.93 -0.21
Red (665nm) 0.73 0.89 -0.16
NIR (842nm) 0.70 0.78 -0.09
SWIR16 (1610nm) 0.054 0.064 -0.010
SWIR22 (2190nm) 0.055 0.067 -0.011

Key observation: Colony has ~20% lower reflectance across all visible bands, but crucially maintains a flat spectral profile. Ice shows characteristic blue-dominated slope.

Discriminating Indices

Index Colony Ice Δ
R-B +0.001 -0.039 +0.040
NIR-R (NDVI) -0.026 -0.064 +0.038
NIR-SWIR 0.856 0.849 +0.007

The R-B index provides the strongest discrimination (+0.040 difference), confirming that RGB-only detection is spectrally sound.

Final Detection Algorithm

Multi-spectral criteria:

colony_mask = (
    (rb_index > -0.01) &      # flat/warm spectrum
    (brightness > 0.65) &      # not shadow
    (brightness < 0.85) &      # not bright ice
    (NIR < 0.75) &             # darker in NIR
    (SWIR16 < 0.065)           # SWIR absorbing
)

Performance comparison:

Method Detection Precision Recall
RGB-only 4.3 ha 31% 76%
Multi-spectral 2.9 ha 33% 54%
7-band detection comparison
7-band detection comparison

Practical Recommendations

  1. For scanning/discovery: Use RGB-only R-B index (high recall)

    candidates = (rb_index > -0.015) & (brightness > 0.6) & (brightness < 0.9)

  2. For validation: Apply multi-spectral criteria (higher precision)

  3. For quantification: Use nested confidence zones

    • Core (R-B > 0): definite guano
    • Extended (R-B > -0.015): likely stained
    • Influence (R-B > -0.025): possibly affected

  4. Change detection: Track area in each zone over time series

Files

Filename Description
colony_map.png 8-bit RGB detection map
colony_final_16bit.png 16-bit nested confidence zones
spectral_signature.png Colony vs ice spectral profiles
spectral_overview.png All 7 bands and derived indices
detection_comparison.png RGB vs SWIR vs combined methods
colony_7band_final.png Final multi-spectral detection

Next Steps

  • Apply to additional known colony sites for threshold calibration
  • Build training dataset with human-labeled regions
  • Test against commercial high-resolution imagery
  • Integrate Sentinel-1 SAR for cloud-free / polar night coverage
  • Automate time series analysis for seasonal monitoring

Analysis conducted using Sentinel-2 L2A data via the estinel pipeline.